Re: NSS issues after locking fixes

From: Alexander Richardson <arichardson_at_freebsd.org>
Date: Mon, 19 Apr 2021 10:06:28 +0100
On Sun, 18 Apr 2021 at 20:58, Александр Недоцуков <bland_at_bbnest.net> wrote:
>
> Hi All,
>
> After supposed locking fixes done in this commit: 5245bf7b92b74e556527b4916a8deba386fe5772
> we have a broken nsdispatch(3) when system has a threaded nss module plugged in.
>
> To illustrate the problem:
>
> $id bland
> id: bland: no such user
> t$LD_PRELOAD=/usr/lib/libthr.so id bland
> uid=1001(bland) gid=500(domainusers) groups=500(domainusers),...
>
> $cat /etc/nsswitch.conf
> group: files ldap
> hosts: files mdns dns
> netgroup: comapt
> networks: files
> passwd: files ldap
> shells: files
> services: files
> protocols: files
> rpc: files
>
> It seems that the commit in question was done under asumption that threading model can never be changed during run-time, while previous code was taking care of it by taking a copy of __isthreaded.
>
> I can not reach out Alex Rechardson who authored this change. Perhaps someone else can step in and take care of it?
>
> Thanks,
> Alexander.

Hi Alexander,

I was away over the weekend so I wasn't able to revert it then. I've
committed the revert as
https://cgit.freebsd.org/src/commit/?id=738314e445ceac4d3dd6c77c636044141623b8dc
and

I am not convinced that __isthreaded changing is the real problem.
This can only happen with fork() (which calls _thr_setthreaded(0))
since if it's currently false, the main thread is executing and we
can't be spawning a new thread.
I think the problem could be that fork() is called while the lock is
held in one thread and therefore the new process ends up with a locked
rwlock. I think the solution could be to add a pthread_atfork()
callback that unlocks the rwlock.

Sorry for the breakage!

Alex
Received on Mon Apr 19 2021 - 07:06:42 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:41:28 UTC